The second method is a structural break test in the GARCH(1,1) volatility

(1)

EU-wide banking stress tests and the effects on equity and CDS markets over time: A comprehensive event study^︎^✩

Master Thesis in Finance, June 2017

Christian Weiss ^a,*, supervised by Dr. M. Hernandez Tinoco ^a

a Faculty of Economics and Business, University of Groningen, The Netherlands

ARTICLE INFO

JEL classification:

E58 E65 G14 G21 G28

Keywords:

Stress tests Event study

Information disclosure Banks

Equity market CDS market

ABSTRACT

EU-wide banking stress tests have become a common supervisory tool to ensure the integrity and financial stability of the European banking system. This paper examines the effects of banking stress tests on stock and credit default swap (CDS) markets over time by applying two different methods. Using the event study methodology, it was found that both markets were not able to anticipate the outcome of stress tests. However, there is some evidence that the stress tests conveyed valuable information to both markets, wherein the reaction on the stock market showed stronger signs than the CDS market. The second method is a structural break test in the GARCH(1,1) volatility. Evidence was found in both markets that banks, which did not do well during the stress tests, were more affected by changes in the variability than banks with sound results. Additionally, the study observes a significant trend in decreasing volatility over time in the stock market, but not in the CDS market, resulting most likely from a generally higher level of confidence by stock investors in the stress test methodology and in the banking industry.

_______________________

✩ The author is profoundly grateful to Dr. M. Hernandez Tinoco for the great guidance and helpful inputs.

* Corresponding author at: University of Groningen, Faculty of Economics and Business, Groningen, The Neth- erlands. Student number: 3066266.

Email: chweiss@live.at (C. Weiss).

(2)

1. Introduction

After the financial crisis in 2008 and the following worldwide distressing spillover effects on financial markets including the alarming capital shortfall of several financial institutions in Europe, the monitoring and steering of risks have become increasingly important, both on the micro and macro level. In order to restore public trust and to ensure the orderly functioning as well as integrity of the financial system in the European Union, the European Banking Au- thority (EBA) conducts EU-wide banking stress tests wherein market developments, trends and potential systemic risks are to be identified. The EBA’s overall aim, together with national supervisory authorities, is the assessment of the resilience of the European financial system against the occurrence of adverse economic shocks.

After 2010, 2011 and 2014 the EBA disclosed the results of the fourth consecutive stress test in July 2016 in which 51 financial institutions were examined, whereas in 2010 91, in 2011 90 and in 2014 123 banks were covered (EBA, 2010, 2011, 2014, 2016). All stress tests are done on a bank-by-bank basis, where banks’ specific data and supervisory information are used. The scope includes the major EU cross-border banking groups and a group of additional, mostly large financial institutions in Europe. In that sense, all stress tests are conducted on a group-consolidated basis including all cross-border operations via branches and subsidiaries.

The stress tests in 2010, 2011 and 2014 were designed to cover at least 50% of the national banking sector of each EU member state and Norway, as expressed in terms of total assets. In the 2016 stress test, the EBA focused more on a homogeneous sample of banks in order to ensure greater comparability by maintaining a significant coverage of EU banking assets. The 51 tested banks in 2016 covered 70% of the Eurozone banking sector including non-Eurozone member state and Norwegian banks with a minimum of EUR 30 billion in terms of total consolidated assets.

Based on the efficient market hypothesis, assets prices are assumed to fully reflect all the information available on the market at any time (Fama, 1970). Provided that markets are efficient and rational, only new, reliable and relevant information should affect prices of securi- ties and other financial instruments. Banking stress tests in general are considered to provide unique information to outsiders about the state of the financial sector (Goldstein and Sapra, 2014). The relation between the disclosure of stress test results and the impact on equity and risk has been part of many studies. For the detection of equity movements, stock prices are used, and for risk, credit default swap spreads (CDS spreads). Both daily stock prices and CDS spreads are part of this study.

(3)

Credit default swaps (CDS) are a popular credit derivative financial instrument that provides insurance against the default by a company, also known as credit event. In the case of a credit event, the buyer of the CDS has the right to sell the bond for the face value to the CDS issuer, also known as reference entity (Hull, 2012). The premium paid for entering a CDS contract is called CDS spread. This spread has a negative relation with the company’s credit rating, meaning that if the company gets downgraded, the CDS spread increases (Hull et al., 2004). As stress test results provide unique information to investors about the financial condi- tion of banks, it can be assumed that this negative relation also holds in case of the event of stress tests. Given the direct influence of the market and the rapidly strong growth of the CDS market over the last decade, CDS spreads are a valuable source for credit risk assessment (Weistroffer et al., 2009).

The objective of this master thesis is divided into two parts. Part one provides a structured overview of the empirical impact of the disclosure of the results of all EU-wide banking stress tests on stock prices and CDS spreads over the period from 2010 to 2016, covering four stress tests overall. The impact on the CDS spreads will be measured indirectly by creating synthetic bonds, a structured combination of a long position in a risk free bond and a short position in the CDS spread. The research question in this respect is whether stress tests provide additional information not only to governments and national supervisory authorities, but also to investors, namely stock and bond investors, and further how this information was evaluated by them. For the first part the event study methodology is applied¹.

The objective of the second part is to give a comprehensive analysis if there is a significant tendency whether the variability of the impact on stock prices and CDS spreads changes over time in terms of volatility, and if so, why. Therefore, the idea behind is to test whether market participants have become either more or less sensitive to new disclosures of stress test results.

To the author’s knowledge, this topic has so far only been discussed in Alves et al. (2015) by examining the 2010 and 2011 stress test. By including the 2014 and 2016 stress test, this study is intended to provide a broader picture and new insights of the impact over time. Even though prior research has been made on the implications of the disclosure of stress test results, new insights about the dimensions of the impact are needed. New insights can be of main interest not only to investors and their investment strategies but also to governments and regu- lators in order to assess the usefulness and effectiveness of banking stress tests with regard to the potential major role of market discipline (Pillar III of the Basel framework) in banking supervision.

1 The entire event study is performed by using the software Event Study Metrics (http://eventstudymetrics.com).

(4)

The study concludes that although both the stock and CDS market were not able to anticipate the outcome of the stress test results in all four stress test exercises, some valuable information was conveyed to both markets after the disclosure of the results. In both markets, some evidence indicates positive market responses for banks that passed the stress tests, un- derlining the achievement of EBA’s goal in decreasing the opaqueness of banks and increas- ing market transparency. On the other hand, only the stock market showed some evidence of negative responses among banks that failed the tests in the form of negative stock returns after the disclosure of the results, whereas the CDS market did not, emphasizing the differences in investor objectives in both markets. The paper further concludes that banks, which did not do well during the four exercises, were more affected by changes in the variability in both direc- tions. Moreover, a trend in decreasing volatility in the stock market before the disclosure of the results was detected, but not in the CDS market, indicating less trading activities by stock investors in the pre-event window. This trend may be caused by various factors: a familiariza- tion by stock investors with the stress testing procedures over the years, a decrease in uncertainty over the years by providing unique information to investors via the disclosure of the results and lastly, a generally higher level of confidence by investors in the banking industry due to better financial conditions of banks in recent years.

The remaining part of this thesis is structured into four main parts. Section 2 outlines the existing literature in this research field. Then, section 3 presents the covered dataset of this thesis. Section 4 introduces the specifications of the event study methodology, parametric and nonparametric significance tests, the structural break test in the GARCH(1,1) volatility and the testable hypotheses. Thereafter, interpretations of the obtained findings are provided in section 5, followed by the conclusion in section 6.

2. Literature

The current body of research with respect to the impact of the disclosure of stress test results in Europe is still scarce, especially about the impact on credit markets. This might come from the fact that CDS markets are compared to other markets relatively young and small (Norden and Weber, 2004). However, CDS markets have grown rapidly in the last decade and, therefore, are able to provide valuable information for risk assessment (Weistroffer et al., 2009). Due to the rise in reliable data for research purposes, research is slowly picking up in this field.

The most comprehensive event study analysis in this field of research for European banks is the 2015 published article by Alves et al. (2015). They conducted an event study and exam-

(5)

ined the impact of the disclosure of EU-wide stress test results from 2010 and 2011 on Euro- pean financial stocks and CDS spreads. They also performed a volatility test before and after the event in order to assess changes in the volatility. Overall, they conclude that stress tests reduce bank opaqueness by conveying new and reliable information to the market. The stock market did not anticipate the outcome of the results, whereas the CDS market partially did.

By creating subsamples of banks based on the fact whether the bank either passed, nearly passed or failed the stress tests, the study concludes further that the outcome was stronger for less capitalized banks, which failed the tests. Although the effect on stock prices showed in both years the same pattern, the effects on CDS spreads were reverse in the 2011 stress test, suggesting two indications: first, both the stock and the credit market show dissimilarity in terms of the ability to incorporate and anticipate new available information. This might come from the fact that CDS market participants are better aware of market developments, whereas equity investors are in general more exposed to market sentiment waves leading to an overre- action in trading behavior. Second, given the nature of equity and debt markets, stock prices reflect the value of the company, whereas CDS market prices reflect the credit risk of the specific bank, including the probability of a credit event. In other words, in case one bank shows a significant capital shortfall after the disclosure of the results and equity investors perceive that the bank needs to raise additional capital, stock prices are likely to adjust accordingly.

But when credit market investors do not believe that this capital shortfall affects the long-term credit risk, CDS spreads remain unchanged. This might explain the reverse reaction in the findings of the 2011 stress test.

Moreover, the results partially confirm the findings of Petrella and Resti (2013) that show that equity market participants considered the information provided in the 2011 EU stress test as important and valuable in order to mitigate bank opacity. However, they also show that the market was not able to anticipate the outcome of the test. Cardinali and Nordmark (2011) show that the 2011 clarification event for the stress test was highly informative, whereas the 2010 results and the release of the 2011 methodology did not contribute valuable information to the market. In contrast, Ellahie (2012) indicates that the disclosure of the 2010 and 2011 results had no significant impact on equity and bid-ask (bond) spreads, but reported a decline after the announcement. Additionally, he shows that though the disclosure of the results reduced information asymmetry, uncertainty deteriorated due to various factors.

Apart from Europe, more research on this topic has been conducted in the US. Morgan et al. (2014) conclude that the disclosure of the 2009 US stress test results reduced the information asymmetry and provided valuable content, though the market already anticipated

(6)

banks with a weaker capitalization. Nevertheless, investors were surprised about the capital gaps of the weaker capitalized banks.

More recently, Neretina et al. (2015) published in a working paper of the Dutch Central Bank the effects of US stress tests on equity prices, credit risk and systemic risk between 2009 and 2015. They conclude that stress tests contribute valuable information for market participants and are a useful tool in mitigating bank opacity. In this study, little was found that stress test outcomes affected banks’ stock prices positively. Only one clarification event by former FED chairman Bernanke in 2009 triggered positive stock market effects. However, at the same time credit markets showed a significant reaction in the form of a decline in CDS spreads in 2009, 2012 and 2010. Furthermore, they also found that the disclosure of stress test results mitigated systemic risk in the years 2009 and 2012 by analyzing the behavior of the banks’ betas (Neretina et al., 2015).

In a different context, Hull et al. (2004) examined the relationship between credit rating announcements, CDS spreads and bond yields based on individual CDS quotes between 1998 and 2002. They conclude that reviews for downgrades affect CDS markets at least 90 days before the actual announcement, but actual downgrades and negative outlooks do not. How- ever, CDS markets anticipated all three types of credit announcements. Another study by Norden and Weber (2004) examined the relationship between credit rating announcements and CDS and stock markets. Their sample included data from credit rating announcements for the period 2000 to 2002. They conclude that both markets forecasted properly rating downgrades and reviews for possible downgrades. Their findings are consistent with Hull et al.

(2004). CDS markets showed earlier reactions than the stock market. They interpret this pattern as such that CDS markets are more efficient than stock markets. This is in line with the findings of Alves et al. (2015) that CDS markets are better aware of market developments compared to stock markets. Lastly, Micu et al. (2006) conclude in their study that all types of rating changes in both ways, negative and positive, have a significant effect on CDS spreads and contain valuable information for credit market participants, although the magnitude of negative rating announcements is much stronger.

3. Data

The EBA examined in the last four stress tests 150 different financial institutions overall from the European Union and Norway, wherein 90, 91, 123 and 51 participated in the 2010, 2011, 2014 and 2016 stress test, respectively. This study includes all four stress tests from 2010 to 2016. The timeline of EU-wide banking stress tests is exhibited in figure 1.

(7)

Figure 1: Timeline of EU-wide banking stress tests²

One part of the dataset includes all the stock data, whereas the other part of the dataset consists of the CDS market spread data. From the 90 participating banks in the 2010 stress test, the data sample includes 49 banks for the stock data sample and 47 for the CDS market spread sample. The variation arises from two facts: first, not all banks are publicly listed, and second, for some banks there is no CDS market spread data available. For the consecutive stress tests, the stock data sample includes 51, 61 and 35 banks, whereas the CDS sample consists of 44, 46 and 32 banks, respectively. For the stock data daily stock prices are used.

The CDS data is based on senior contracts with a maturity of 5 years, which is the standard for CDS market studies. Furthermore, all CDS contracts are denominated in Euro and almost all are restructured with the preferred type “Modified-Modified” (MM)³, as this type represents the standard of European CDS contracts. Table A1 in the Appendix exhibits the full list of all participating banks of the four stress tests and outlines further in which data samples the banks are part of.

By including banks whose operations were discontinued due to whatever reasons, the sur- vivorship bias is eliminated, which should lead to more robust conclusions (Elton et al., 1996).

In order to strengthen the results of the study, a control group has been added for both the stock and the CDS market. The control group consists of listed financial institutions and financial companies that have not participated in one of the four stress tests (Table A2 and table A3 in Appendix).

To detect abnormal performances, benchmark indices are required. The benchmark indices should represent normal returns of the industry in which the examined companies are operat-

2 As the 26/10/2014 was not a trading day (Sunday), the 27/10/2014 represents the event day in this study.

3 Since not all banks have MM structured CDS contracts, the fully restructured (CR) or non-specified CDS contract (non-spec.) is used in that case. The CDS contract type of each bank is shown in table A1 in the Appendix.

23/07/2010 Disclosure of 2010 stress test

26/10/2014 ² Disclosure of 2014 stress test

29/07/2016 Disclosure of 2016 stress test 15/07/2011

Disclosure of 2011 stress test

(8)

ing. The reference index for the stock market sample is a European banking price index from Datastream (Datastream Symbol: BANKSEU) representing the performance of 125 European banks. For the CDS market sample the Markit iTraxx Europe Senior Financial Index (Bloom- berg⁴: SNRFIN CDSI GEN DY Corp) is used as benchmark index. This index comprises 30 equally weighted CDS contracts with maturities of 5 years on investment grade European entities and matches fully the restructuring type of the underlying data sample.

Based on whether banks passed or failed the stress test, the data sample is divided into subsamples. This procedure is in line with Alves et al. (2015). The results from the stress tests are retrieved from the EBA’s website. Group A consists of banks that passed the stress test, group B are banks with tangential⁵ results and banks that failed the test are in group C. Table 1 exhibits the critical values for the group differentiation based on the adverse Common Equity Tier 1 Ratio (CET1 Ratio) in the publication of the stress test results published by the EBA. In the 2016 stress test the EBA removed the definition of a specific capital hurdle from the publication. They explained this step by referring to the strong increase of the average CET1 ratios of the participating banks within the last 6 years and by moving from an “capital now”

approach to a “forward looking capital planning” approach (EBA, 2016). Furthermore, this new approach enables market participants to make their own assessments with regard to market discipline. Thus, the capital hurdle of 6.5% for the 2016 stress test is set based on the improved capital requirements of the average bank subject to the stress test. In which subsample one particular bank of the dataset is allocated, is displayed in table A1 in the Appendix.

Table 1: Thresholds for subsamples based on adverse CET1 Ratio

Group A Group B Group C

Passed Tangential results Failed

2010 Stress test* > 6.0% ≤ 5.0% to ≤ 6.0% < 5.0%

2011 Stress test* > 6.0% ≤ 5.0% to ≤ 6.0% < 5.0%

2014 Stress test** > 6.5% ≤ 5.5% to ≤ 6.5% < 5.5%

2016 Stress test** > 7.5% ≤ 6.5% to ≤ 7.5% < 6.5%

Notes:

* CET1 Ratio based on two year adverse scenario

** CET1 Ratio based on three year adverse scenario

4 I would like to profoundly thank Ivan Soldo, my former team head at Raiffeisen Bank International AG, for providing me access to Bloomberg.

5 Tangential means that the particular bank nearly passed the stress test, but was also very close to failing.

(9)

4. Methodology

4.1. Event study methodology

For the first part of this study the event study methodology is used to examine stock market and CDS market behavior around the disclosure date of the stress test results. The methodology applied is derived from the general event study methodology as discussed in MacKinlay (1997). Thus, the estimation window includes 120 trading days ([t-129; t-10]) prior to the event date ([t0]). For the so-called event window five different alternatives are consid- ered in order to fully cover the entire range of possible empirical impacts: (i) [t-9; t-1]; (ii) [t-5; t-1]; (iii) [t-1; t1]; (iv) [t0; t5]; (v) [t0; t10]. For a detailed discussion of the event study methodology and its historical development see MacKinlay (1997).

Figure 2: Timeline event study

As mentioned in the introduction, the impact on the CDS market is measured indirectly by creating synthetic bond yields⁶. This structuring method enables replicating the bond’s cash streams and is in line with the suggestion by Hull et al. (2004), who show that this theory holds. The synthetic bond yield is structured by taking a long position on a risk-free par bond⁷ with a maturity of 5 years and a short position on the respective 5-year CDS contract. The outcome is identical to holding an actual bond of the particular bank, but without any direct involvement of the bank in the transaction, which is useful due to the occasional constraints in bond markets. Considering this, the method is a simple way to homogenize the results for comparing the effects of stress tests on a larger scale. Another benefit of the method is that interest rate risk and counterparty default risk are combined, covering overall two major risk factors which are both of most interest to investors.

Synthetic bond yield = Risk-free yield on a par bond + CDS spread

(Long position) (Short position)

6 For the remainder of this study, the definition “synthetic bond“ represents CDS spread.

7 Given the fact that this study examines EU-wide banking stress tests, the 5-year German Bund yield is used.

t_-129 t_-10 t_-9 t₀ t10

Estimation window Event window

(10)

To measure whether the disclosure of stress test results has an effect on stock and CDS prices, abnormal returns (AR) are calculated. The basis for computing normal returns is the market model, as this model is an improvement over the constant mean return model in terms of reducing the variance of the AR (MacKinlay, 1997). The equation for the market model is as follows:

!_!" = !_! + !_!!_!" + !_!" (1)

where !_!" is the return of security i (synthetic bond i) in t, !_! is the intercept of the return at time t, !_! is the covariance of the return of security i (synthetic bond i) with the market portfo- lio, !_!" is the return on the market portfolio and !_!" is the error term of security i (synthetic bond i) in t. The model parameters !_! and !_! are estimated over the estimation windows of 120 days by ordinary least squares regressions (OLS). The market model is derived from the capital asset pricing model (CAPM) and assumes a risk free rate of zero. For both data samples, the returns are computed by using continuously compounded returns:

!_!" = ln !_!"

!_!,!!! (2)

where !_!" is the continuously compounded return of security i (synthetic bond i), !_! is the price at t and !_!!! is the price at t-1. Based on the market model, one can calculate the AR as follows:

!"_!" = !_!"−(!_! + !_!!_!") (3)

By aggregating the AR over different event windows, one can see how AR develop over time. The equation for cumulative abnormal returns (CAR) is as follows:

!"#_!(!_!, !_!) = !"_!"

!_!

!!!!

(4)

where !"#_!(!_!, !_!) defines the cumulative average return (CAR) of security i (synthetic bond i) in the respective event window. For a greater comparability of the different event windows, the average abnormal return (AAR) and the cumulative average abnormal returns (CAAR) are calculated:

(11)

!!"_! = 1

! !"_!"

!

!!!

(5)

!""#(!_!, !_!) = !!"_!"

!_!

!!!!

(6)

or

!""#(!_!, !_!) = 1

! !"#_!(!_!, !_!)

!

!!!

(7)

For simplicity, in the remainder of this study, the event study terminology is employed when referring to abnormal changes in the stocks prices (CDS spreads) as AR in stock prices (CDS spreads).

4.2. Significance tests

The event study includes three parametric and two nonparametric tests. The student t-test, the cross-sectional t-test and the adjusted standardized cross-sectional test by Boehmer et al.

(1991) with the extension by Kolari and Pynnönen (2010) are the three parametric tests. The Corrado rank test (Corrado, 1989) and the Cowan’s generalized sign test (Cowan, 1992) represent the two nonparametric tests. Traditional parametric tests assume that AR are cross- sectionally uncorrelated, but due to clustering this is not valid when firms are from the same industry (see Collins and Dent 1984; Brown and Warner, 1980, 1985; Kothari and Warner 2007). Another disadvantage is that traditional parametric tests are prone to event-induced volatility, as discussed in other studies (see Collins and Dent 1984; Brown and Warner 1980, 1985; Kothari and Warner 2007; Harrington and Shrider, 2007). Additionally, parametric tests assume that returns are normally distributed. However, Brown and Warner (1985) conclude that this assumption of normality is violated in case of security returns. All downsides of parametric tests can lead to a misspecification of test statistics and can, consequently, cause sub- stantial over-rejections of true null hypotheses. Thus, by including the adjusted standardized cross-sectional test and nonparametric tests, the chance of detecting false null hypotheses increases. In particular, nonparametric tests strengthen the robustness of the results (MacKinlay, 1997).

For a detailed review of parametric and nonparametric event study tests see Dutta (2014).

(12)

4.2.1. Student t-test

The student t-test is the most common test in detecting significant abnormal returns (CAAR). Brown and Warner (1980) conclude that this test has high power, but only if strict conditions are met. One of the main assumptions is the normality in security returns. To test whether this assumption is violated, the study includes a normality test in the form of the Jarque-Bera test statistic (Table 2). Further downsides of the student t-test are the proneness to cross-sectional correlation and to changes in the volatility. The student t-test is defined as follows:

!_!" = !""#_!

(!_!− !_!+ 1)^!^! !_!!"_! (8)

The variance of this statistic is based on the time-series of AR from the estimation window:

!^!_!!"_! = 1

! − ! !!"_!− 1

! !!"_!

!"#_!"#

!"#!"#

!"#_!"#

!!!"#!"#

!

(9)

where M is the number of non-missing returns and d the degrees of freedom, which is in the case of the market model ! = 2. Due to the fact that the event window AR are an out of sample prediction, the standard error is adjusted by the forecast error (Event Study Metrics, 2011).

Thus, the adjustment in the market model is as follows:

1 + 1

!+ (!_!" − !_!"#$)^!

(!_!" − !_!"#$)^!

!"#_!"#

(10)

The test hypothesis is: H0: !""#(!_!, !_!) = 0 H1: !""!(!_!, !_!) ≠ 0

4.2.2. Cross-sectional t-test

The cross-sectional t-test is a simple extension of the student t-test by dividing the CAAR by the cross-sectional standard error. The test is defined as:

!_!"#$$ = !""#(!_!, !_!)

!!""#_(!!,!!) (11)

(13)

The cross-sectional variance is given as:

!^!!""#_(!!,!!) = 1

!(! − !) [!"#_!(!_!, !_!) − !""#(!_!, !_!)]^!

!

!!!

(12)

where N is the number of observations and d the degrees of freedom. The test hypothesis is the same as in the student t-test. Brown and Warner (1985) show that the cross-sectional t-test is prone to event-induced volatility. Therefore, this method might reject the null hypothesis too frequently. Same as the student t-test, the cross-sectional t-test assumes normality in the return distribution and is also prone to cross-correlation of security residuals.

4.2.3. Adjusted standardized cross-sectional t-test

Boehmer et al. (1991) proposed a standardized cross-sectional parametric method, which is robust to event-induced variance increases of stock returns. The cross-sectional average return (CSAR) is defined as:

!"#$(!_!, !_!) = 1

! !"#$_!(!_!, !_!)

!

!!!

(13)

The standard deviation of !"#$(!_!, !_!) is estimated from the cross-section of event- window AR:

!(!"#$) = 1

!(! − 1) [!"#$_!(!_!, !_!) − !"#$(!_!, !_!)]^!

!

!!!

(14)

The standardized cross-sectional test statistic for the null hypothesis that the CAAR is equal to zero is denoted as follows:

!!"#!!"# !" !".= !"#$(!_!, !_!)

!(!"#$) (15)

The main downside of the standardized cross-sectional test is that the test is not robust to cross-correlation of security residuals. Therefore, Kolari and Pynnönen (2010) proposed an adjustment to the test to also account for cross-correlation. The adjusted standardized cross- sectional test statistic for the null hypothesis that the CAAR is equal to zero is:

(14)

!!"#!!"# !" !".,!"#. = !!"#!!"# !" !".

1 − !

1 + (! − 1)! (16)

where ! denotes the average cross-correlation among AR. According to Kolari and Pynnönen (2010) the adjusted standardized cross-sectional test shares the robustness and power proper- ties with the rank test in shorter event windows and even outperforms the rank test in longer event windows.

4.2.4. Corrado rank test

Given the fact that security returns are not normally distributed, the assumption of normality is violated (Brown and Warner, 1985). Thus, two nonparametric tests, namely the rank and the generalized sign test, are included. Both tests do not require symmetry of the cross- sectional AR distribution (Dutta, 2014). The first is the Corrado rank test, which tests the null hypothesis that the AAR is equal to zero (Corrado, 1989). This test transforms the AR into ranks, asset by asset for the joint time period including estimation window and event window:

!_!,! = !"#$(!"_!,!) (17)

Corrado and Zivney (1992) suggest a uniform transformation of ranks to adjust for missing values:

!_!,! = !_!,!

(1 + !_!) (18)

where !_! is the number of non-missing returns for each asset. The single day test statistic is then defined as:

!!"##$%! = 1

! (!_!,!− 0.5)/!(!) (19)

The standard deviation is estimated as follows:

!(!) = 1

!_!+ !_!

1

!_! (!_!,!− 0.5)

!_!

!!!

!

(20)

(15)

where !_! is the number of non-missing cross-sectional returns at ! = !. The multiday standard deviation is estimated by taking the average of single day statistics multiplied by the in- verse of the square root of the period’s length. The rank test is, when correctly specified, more powerful in shorter event windows than the generalized sign test. Nevertheless, the rank test suffers from a deteriorated performance in case of increased variance, causing over-rejections of true null hypotheses (Cowan, 1992; Cowan and Sergeant, 1996).

4.2.5. Cowan’s generalized sign test

The second nonparametric test is the generalized sign test, which was proposed by Cowan (1992). The test is based on the ratio of positive CAR !_!^! over the event window. Under the null hypothesis this ratio should be in line with the ratio of positive CAR over the estimation window !_!"#^! . The ratio of positive CAR is a binomial random variable. The following test statistic is used:

!_!" = !_!^!− !_!"#^!

!_!"#^! (1 − !_!"#^! )/! (21)

The test statistic follows a normal distribution. Besides that the test does not require symmetry in the return distribution, other advantages, also relative to the rank test, are a better performance under increased variance and superior results in longer event windows (see Cowan, 1992; Cowan and Sergeant, 1996; Campbell et al., 2010).

4.3. Volatility test

The second part of this study examines volatility changes before, during and after the disclosure of stress test result. The methodology applied is based on a GARCH(1,1) model⁸, which is defined as:

!_!" = !_!+ !_!!_!" + !_!

!_!^! = !_!+ !_!!_!!!^! + !_!!_!!!^! + ∅ !" (22)

where !_!" is the return of security i in t; !_!" is the return on the market portfolio in t; !" is a binary variable (dummy variable) equal to either 1 in the event window and 0 in all other observations, !_! is the conditional variance and !^! is the unconditional variance. In case of a

8 To minimize the failure rate in the calculations, stock-, synthetic bond- and index returns have been standardized, ranging from 0 to 1, where 0 corresponds to the minimum and 1 to the maximum.

(16)

statistically significant positive (negative) ∅, volatility increased (decreased) in the event window.

4.4. Hypotheses

The study examines the impact of EU-wide banking stress tests on stock and CDS markets over time. The following hypotheses are in line with prior research.

Harrington (2006) shows that CDS markets are able to anticipate market developments due to information leakages, market expectations or speculation. Therefore, if markets anticipate the results of the stress tests, prices start to adjust already before the publication. In this case, the event windows [t-9; t-1] and [t-5; t-1] are examined. The hypothesis is as follows:

H1: Both stock prices and CDS spreads anticipate the direction of EBA’s stress test outcomes.

The disclosure of banking stress tests results convey valuable information in the form of capital adequacy, risk appetite and asset quality to outsiders and might lead to an adjustment in the assessment of the participating banks. Since stock prices reflect the value of the bank, the value might change after the disclosure of new private information. Contrary to the stock market, the new available information might change the credit risk assessment, which can be seen in the form of adjusted CDS spreads. Therefore, the hypothesis is as follows:

H2: The publication of the stress test outcomes has informational content for both the stock and the CDS market.

The study includes a control group in all test samples for both markets in order to provide more robustness and to evaluate whether stress tested banks performed differently than the control group. The control group consists of banks and financial companies, which have not been part of any stress tests (Table A2 and A3 in the Appendix). In this respect, the study tests the following hypothesis:

H3: The impact of the disclosure of the stress test results is higher for participating banks than for non-participating banks and financial companies.

Good news about the company’s value affects both the value of the company’s common equity and the value of debt in a positive way (Merton, 1974). Therefore, this study assumes that in case one bank passes the stress test, stock prices increase and CDS spreads decrease and vice versa for banks that failed the stress test, leading to the following hypothesis:

(17)

H4: CDS spread changes and stock return changes move inversely. Banks that pass the stress test (Group A) exhibit positive stock performance and negative CDS spread changes. Banks that failed the stress test (Group C) exhibit negative stock performance and positive CDS spread changes.

After four stress tests, this study examines in a last step whether the variability of the impact on stock prices and CDS spreads changes over time in terms of volatility. Banks have improved their core capital ratios substantially over the recent years and economic growth is slowly picking up, resulting in a better asset quality of banks. Thus, it can be assumed that the variability of stock prices and CDS spreads decreases over time. Generally, investors dislike times of high volatility (Ang, 2014). Therefore, it is particularly interesting how volatility develops in the event of banking stress tests. The testable hypotheses are as follows:

H5: The change in the volatility in both the stock prices and CDS spreads is greater for banks that failed or nearly passed the stress tests than for banks that passed the stress tests.

H6: Provided that stress tests convey new and valuable information to outsiders, volatility decreases in the pre-event windows and increases after the disclosure of the results in both markets in the post-event windows.

H7: Stress testing as a supervisory tool decreases the volatility in both the stock market and CDS market over time.

5. Results

5.1. Event study results

Table 2 exhibits the descriptive statistics for the AR in the estimation windows for all of the four stress tests for both the stock market sample and the CDS spread sample in the form of synthetic bonds. Furthermore, the respective control groups are included. The descriptive statistic includes 120 trading days for each of the sample. The mean of the AR is zero in all data samples. The standard deviations of the synthetic bonds show higher results compared to the ones of stock prices. To test whether the assumption of linearity in the AR for the different time samples is violated, the kurtosis, skewness and Jarque-Bera (JB) statistics are included.

Except for the 2011 and 2014 stock data sample, as well as the 2014 stock control group sample, the AR are not normally distributed, as they show significant JB test statistics above 9.2.

(18)

Furthermore, all synthetic bond samples exhibit JB test statistics beyond 9.2. In all those cases, the null hypothesis that the stock (synthetic bond) returns are normally distributed is rejected at the 1% level. These results clearly justify the application of nonparametric significance tests, which do not require symmetry of the cross-sectional return distribution.

Table 2: Descriptive statistics of abnormal returns of stocks and synthetic bonds in estimation windows

²⁰¹⁰ ²⁰¹¹ ²⁰¹⁴ ²⁰¹⁶ ^Control₂₀₁₀ ^Control₂₀₁₁ ^Control₂₀₁₄ ^Control₂₀₁₆

Panel A - Stocks

^Mean 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

^Median -0.0004 0.0001 -0.0002 0.0003 0.0007 0.0001 0.0000 0.0003

Standard deviation 0.0059 0.0058 0.0045 0.0073 0.0052 0.0039 0.0035 0.0039

^Minimum -0.0121 -0.0199 -0.0102 -0.0279 -0.0182 -0.0166 -0.0079 -0.0135

^Maximum 0.0232 0.0137 0.0144 0.0217 0.0228 0.0135 0.0116 0.0130

^Kurtosis 4.0509 3.7294 3.0762 5.2224 6.8984 7.4820 3.6328 4.4879

^Skewness 0.5198 -0.3582 0.1976 -0.5208 0.3870 -0.6040 0.3456 -0.4553

Jarque-Bera (JB) 10.9265 5.2260 0.8102 30.1183 78.9805 107.7395 4.3907 15.2149

Estimation window (T) 120 120 120 120 120 120 120 120

Panel B - Synthetic bonds

^Mean 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

^Median 0.0000 -0.0002 0.0002 -0.0003 -0.0007 -0.0005 -0.0003 -0.0008

Standard deviation 0.0157 0.0109 0.0100 0.0158 0.0184 0.0121 0.0145 0.0228

^Minimum -0.0511 -0.0348 -0.0399 -0.0576 -0.0648 -0.0306 -0.0675 -0.0750

^Maximum 0.0502 0.0400 0.0234 0.0877 0.0778 0.0531 0.0428 0.1123

^Kurtosis 4.8056 4.5303 5.7622 14.8484 5.7963 6.3101 6.7919 8.2291

^Skewness -0.2812 0.3466 -0.9362 1.9105 0.4194 1.0488 -0.5581 0.9110

Jarque-Bera (JB) 17.8812 14.1109 55.6798 774.9239 42.6149 76.7842 78.1217 153.3178

Estimation window (T) 120 120 120 120 120 120 120 120

Notes: The critical values for the JB test statistic are 6 and 9.2 at the 5% and 1% significance level, respectively.

The results of the event study are displayed in tables 3 to 6 and tables A4 to A5 in the Appen- dix, wherein table 3 shows the percentage of statistically significant CAAR (p-value < 5%) in stocks (Panel A) and CDS spreads (Panel B), and tables 4 to 6 and tables A4 to A5 in the Ap- pendix exhibit the results of the different significance tests, including three parametric and two nonparametric tests. All results of the aforementioned tables are structured by panel and by group of banks (i.e. treatment group, group A, group B, group C and control group) and are presented in a pooled sample, as well as separately for the 2010, 2011, 2014 and 2016 stress test. For further illustration, figures 2 to 6 displays the average CAAR (in %) over the entire event period [t-9; t10] for the pooled data.

Hypothesis 1 (H1) asks if the stock and the CDS market anticipate the outcome of the stress tests. The event windows of interest in this case are [t-9; t-1] and [t-5; t-1]. As seen in table 3, the percentage of statistically significant CAAR at the 5% level based on the student t- test exhibit similar patterns in both the stock and the synthetic bond sample. In the 2010 stress test neither the stock market nor the CDS market anticipated the outcome of the test. The per-

(19)

centage of statistically significant CAAR on a single basis (Table 3) is close to zero in both samples. Only group A in the stock sample shows significantly positive CAAR in the [t-9; t-1] event window. However, the generalized sign test supports this pattern on average, even though the CAAR are in this case negative (Table 6). By further comparing the different significance tests for the 2010 run-up event windows, one can see that both the cross-sectional t- test and the generalized sign test display negative CAAR at the 1% level for group A in the synthetic bond sample, leading to a decrease in the CDS spreads (Table A4 in Appendix).

This is the only case in which the CDS market anticipated the outcome of the stress test results. The results of the student t-test (Table 4), the adjusted standardized cross-sectional (Ta- ble 5) and the rank test (Table A5 in Appendix) display on average no significant CAAR for the 2010 run-up event windows. The control groups show a similar performance in both markets prior to the event.

In case of the 2011 stress test, the percentage of statistically significant CAAR on a single basis (Table 3) show significant results in both markets for the event windows prior to the disclosure. Group B in the stock data sample tend to have negative CAAR on average, which is supported by 5 out of the 6 significance tests, except for the adjusted standardized cross- sectional test (Table 5). For group C this pattern can only be confirmed by the student t-test.

On the other hand, almost one third of the banks in group A and B display significantly positive CAAR prior to the event on a single basis (Table 3), resulting in an increase of the CDS spreads. This performance can be verified by both the student t-test (Table 4) and the cross- sectional t-test (Table A4 in Appendix) and only for group A by the generalized sign test (Ta- ble 6). The results of the control group did not support any statistical impact in both markets.

According to the results of the 2014 stress test, the evidence for anticipation is weak and not robust. Only the average CAAR of group A in the stock sample reflect positive signs at the 5% level, but this is only supported by the student t-test (Table 4) and the cross-sectional t-test (Table A4 in Appendix). All other tests do not share this behavior. At the same time, the results of the synthetic bond sample indicate a decrease in spreads for group A and partially for group B, but only for the shorter pre-event window [t-5; t-1]. The control group of the synthetic bond sample supports the performance in both markets.

With regard to the latest stress test in 2016, the percentage of statistically significant CAAR on a single basis shows no impact at all (Table 4). Nevertheless, by taking into account average significant CAAR, the significance tests indicate, indeed, an impact in both markets. Group C in the stock sample results reflects in the shorter pre-event window [t-5; t-1] significantly negative CAAR, as shown by the cross-sectional test (Table A4 in Appendix)

(20)

and the adjusted standardized cross-sectional test (Table 5). All other test statistics do not signal any evidence of anticipation. The results in the synthetic bond sample are dissimilar to that, as group C displays negative CAAR in the longer pre-event window [t-10; t-1], indicating a decrease in the risk assessment by investors, which is not in line with the outcome of the stress test. This finding is confirmed by the student t-test (Table 4) and by the rank test (Table A5 in Appendix), a nonparametric test. Additionally, it seems that investors in credit markets might have anticipated the outcome for group A, as both the cross-sectional t-test (Table A4 in Appendix) and the generalized sign test (Table 6) show significantly negative CAAR.

However, considering similar results in the control group undermines the findings, as the impact might be driven by other factors.

In light of the above, there is little evidence overall that both the stock and CDS market anticipated the outcome of the results. This also confirms the results of the pooled samples in both markets. Thus, H1 is rejected and the study concludes that both markets were not able to anticipate the outcome of stress tests.

Hypothesis 2 (H2) claims that the publication of the stress test results has informational content for both the stock and CDS market. The event windows of interest for testing H2 are [t0; t5] and [t0; t10]. At first, by looking at the percentage of significant CAAR based on the student t-test (Table 3) of the pooled stock sample, 7.14% of the banks show increased CAAR in the smaller post-event window [t0; t5]. This result can be mainly attributed to group A. On the other hand 17.39% of failing banks (Group C) exhibit significantly negative CAAR for the same window and 8.70% for the longer one. On an aggregated level, these findings are confirmed by the student t-test (Table 4) and partially by the cross-sectional t-test (Table A4 in Appendix). However, the more robust test statistics only present scant evidence for this.

However, by splitting up the results for the single stress tests, one can see that in contrast to the 2010 and 2016 stress tests only the 2011 and 2014 exercises might have transmitted valuable information to the stock market. In fact, the generalized sign test (Table 6) reflects similar positive CAAR for group A and negative CAAR for group C in the 2014 stock sample as the student t-test (Table 4) and the cross-sectional t-test results (Table A4 in Appendix). The results of the rank test (Table A5 in Appendix) confirm the findings for group A that the 2014 stress test added informational value to the stock market. Overall, the impact on banks’ stock returns that passed the stress test can only be seen in the smaller post-event window [t0; t5], whereas the negative effect on banks that failed the test continued over the longer period of 10 days. Important to mention in this respect is that after the 2011 stress test results, group A and B reflect significantly positive CAAR in the 5 days after the disclosure. This is confirmed